TL;DR
This paper introduces a novel wearable hand module using Fiber Bragg Grating (FBG) strain sensors that can adapt to different hand sizes for accurate real-time finger joint angle tracking. They developed and evaluated four algorithms (PTS, AREA, PS, PS/I_α), finding that PS and PS/I_α were optimal for index-little fingers, and PS/I_α and AREA for thumbs, achieving an average error angle of just 0.47 ± 2.51°.
What Problem Does This Paper Address?
Accurate measurement of hand and finger movements is crucial for applications like rehabilitation, virtual/augmented reality, and robotics. Existing methods, such as optical tracking, are often expensive and limited by space, while Inertial Measurement Units (IMUs) suffer from drift and complexity. Strain sensors offer a promising alternative, but traditional glove-based systems struggle with accuracy across different hand sizes because fixed sensors don't account for variations in joint positions and finger lengths. This leads to inconsistent performance and requires recalibration for each user. This paper aims to overcome these limitations by developing a flexible FBG sensor system and adaptive algorithms that can maintain high accuracy across a range of hand sizes without user-specific recalibration.
Methods
The core of this work is a wearable hand module incorporating five Fiber Bragg Grating (FBG) strain sensors. Unlike fixed sensors, these FBGs are embedded in a flexible TPU guide, allowing them to move continuously with the finger's bending and adapt to different hand sizes. This mobility is key to their approach.
To interpret the FBG sensor data, four algorithms were proposed and evaluated:
- Point Strain (PTS): A basic method using data from a single FBG node per joint.
- Area Summation (AREA): Uses data from multiple FBG nodes per joint, summing their bending percentages.
- Proportional Summation (PS): Weights the contribution of each FBG node based on its strain, giving less weight to nodes that might be dislodged.
- Proportional Summation/Interference (PS/I and PS/I_α): These are advanced versions of PS that specifically address interference from adjacent joints (like the PIP joint affecting the MCP joint measurement). PS/I_α is particularly designed to account for how the influence of a joint's bending percentage changes with hand size.
The researchers also developed a 3D printed hand replica system to quantitatively evaluate the module's performance across various finger lengths, allowing for controlled and repeatable measurements.
Datasets
The paper does not describe training datasets. Instead, it focuses on evaluating the performance of the wearable hand module using a custom-built 3D printed hand replica. This replica was designed to simulate different finger sizes, specifically:
- Index-little fingers: Lengths ranging from 77 mm to 117 mm.
- Thumb: Lengths ranging from 68 mm to 78 mm.
The quantitative evaluation involved measuring joint angles at fixed positions (0°, 20°, 40°, 60°, 80°, 90°) and comparing the module's readings against goniometer measurements.
Key Insights and Ideas
- Sensor Mobility is Crucial: The paper strongly emphasizes that the ability of FBG sensors to move and adapt to finger bending, rather than being rigidly fixed, is fundamental to achieving high accuracy across different hand sizes.
- Algorithm Adaptability is Key for Complex Joints: For joints like the Metacarpophalangeal (MCP) joint, where sensor node positions can vary significantly with hand size and finger movement, simple algorithms like PTS are insufficient. More sophisticated algorithms like PS/I_α are necessary to account for these positional changes and interference from other joints.
- Hand Size Dictates Optimal Algorithm: The "best" algorithm isn't universal. The study found that for index-little fingers, PS and PS/I_α were optimal, while for the thumb, PS/I_α and AREA performed best. This highlights the need for adaptive or size-aware algorithms.
- FBG Sensors Offer High Linearity: The FBG sensors themselves demonstrated excellent linearity and durability, even after 10,000 bending cycles, providing a robust foundation for accurate measurements.
Key Results
The wearable hand module, utilizing FBG strain sensors and optimized algorithms, demonstrated high accuracy in measuring finger joint angles across different hand sizes:
- Overall Performance: The average error angle for all joints, sizes, and conditions was 0.47 ± 2.51°, with a Mean Absolute Error (MAE) of 1.63 ± 1.97°.
- Index-Little Fingers:
- DIP/PIP joints: PS algorithm yielded an angle error of 0.82 ± 2.71°.
- MCP joint: PS/I_α algorithm achieved a significantly reduced error of -0.21 ± 2.32°.
- Thumb:
- IP joint: PS/I_α algorithm showed an error of 0.74 ± 2.84° (large thumb) and -0.35 ± 1.23° (small thumb).
- MCP joint: PS/I_α algorithm achieved an error of -0.05 ± 1.79° (small thumb) and 2.06 ± 2.53° (large thumb).
- Maximum Angle Error: The largest maximum angle error observed was 4.59° for the thumb's MCP joint.
- Linearity: The system achieved high linearity (R² values around 0.998) when using the optimized algorithms, approaching the linearity of the bare FBG sensors.
Limitations and Future Work
- Limited Hand Size Range: The study primarily focused on adult male hand sizes. The authors acknowledge that the gap between FBG nodes needs to be further reduced, and corresponding algorithms developed to accommodate the hand sizes of adult females and adolescents.
- Flexion/Extension Focus: The current work primarily measures flexion and extension. Future research should aim for exact measurements of abduction/adduction and hyperextension movements.
- Algorithm Selection: While a flowchart for algorithm selection is provided, further work could focus on developing a single, adaptive algorithm that automatically adjusts based on hand size and joint characteristics, rather than requiring manual selection.
- Real-World Application: While the 3D printed replica provided controlled evaluation, future work would involve testing the module on actual users in various application scenarios (rehabilitation, VR/AR, robotics) to validate its performance in dynamic, real-world conditions.
Summary updated